org.nd4j.autodiff.samediff.SDVariable Maven / Gradle / Ivy
The newest version!
/*
* ******************************************************************************
* *
* *
* * This program and the accompanying materials are made available under the
* * terms of the Apache License, Version 2.0 which is available at
* * https://www.apache.org/licenses/LICENSE-2.0.
* *
* * See the NOTICE file distributed with this work for additional
* * information regarding copyright ownership.
* * Unless required by applicable law or agreed to in writing, software
* * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
* * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
* * License for the specific language governing permissions and limitations
* * under the License.
* *
* * SPDX-License-Identifier: Apache-2.0
* *****************************************************************************
*/
package org.nd4j.autodiff.samediff;
import lombok.*;
import lombok.extern.slf4j.Slf4j;
import org.nd4j.autodiff.functions.DifferentialFunction;
import org.nd4j.autodiff.samediff.internal.SameDiffOp;
import org.nd4j.autodiff.samediff.internal.Variable;
import org.nd4j.common.base.Preconditions;
import org.nd4j.linalg.api.blas.params.MMulTranspose;
import org.nd4j.linalg.api.buffer.DataType;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.linalg.api.ops.impl.shape.CreateView;
import org.nd4j.linalg.api.shape.LongShapeDescriptor;
import org.nd4j.common.util.ArrayUtil;
import org.nd4j.linalg.factory.Nd4j;
import org.nd4j.weightinit.WeightInitScheme;
import java.io.Serializable;
import java.util.ArrayList;
import java.util.Arrays;
import java.util.Map;
@Data
@NoArgsConstructor
@Slf4j
public class SDVariable implements Serializable {
protected SameDiff sameDiff;
@Getter
protected String varName;
@Getter
@Setter
protected VariableType variableType;
@Setter(AccessLevel.NONE)
protected long[] shape;
@Getter (AccessLevel.NONE)
@Setter
protected DataType dataType;
private DifferentialFunction creator;
// autogen_tag::sdvars::start
public SDVariable(@NonNull String varName, @NonNull VariableType varType, @NonNull SameDiff sameDiff, long[] shape, DataType dataType){
if(varType != VariableType.PLACEHOLDER)
Preconditions.checkState(dataType != DataType.UNKNOWN, "Unknown datatype is not allowed for SDVariables (variable name: %s)", varName);
if(varName == null)
varName = sameDiff.generateNewVarName(varName, 0, true);
this.sameDiff = sameDiff;
this.varName = varName;
this.variableType = varType;
this.dataType = dataType;
this.shape = shape;
}
/**
* Get the name of the SDVariable
* @return Name of the variable
*/
public String name(){
return varName;
}
public void setVarName(String varName) {
this.varName = varName;
}
/**
* @deprecated Use {@link #name()}
*/
@Deprecated
public String getVarName(){
return name();
}
/**
* Returns true if this variable is a placeholder
* @return
*/
public boolean isPlaceHolder() {
return variableType == VariableType.PLACEHOLDER;
}
public boolean isConstant(){
return variableType == VariableType.CONSTANT;
}
/**
* A getter for the allocated ndarray with this {@link SDVariable}.
*
* This getter will lazy initialize an array if one is not found based on the associated shape and
* {@link WeightInitScheme} - if this is possible. If this is not possible (due to shapes being unknown, etc)
* null is returned
*
* @return the {@link INDArray} associated with this variable.
*/
public INDArray getArr() {
return getArr(false);
}
// autogen_tag::sdvars::end
/**
* A getter for the allocated ndarray with this {@link SDVariable}.
*
* This getter will lazy initialize an array if one is not found based on the associated shape and
* {@link WeightInitScheme} - if this is possible.
* If this is not possible (due to shapes being unknown, etc) either:
* (a) null is returned - if enforceExistence == false, or
* (b) an IllegalStateException is thrown, if enforceExistence == true
*
* @return the {@link INDArray} associated with this variable.
*/
public INDArray getArr(boolean enforceExistence) {
if(sameDiff.arrayAlreadyExistsForVarName(getVarName()))
return sameDiff.getArrForVarName(getVarName());
if(variableType == VariableType.ARRAY && enforceExistence) {
throw new UnsupportedOperationException("Cannot get array for ARRAY type SDVariable - use SDVariable.exec or SameDiff.output instead");
} else if(variableType == VariableType.ARRAY) {
if(sameDiff.isEagerMode()) {
return sameDiff.getEagerArrForVarName(name());
}
return null;
}
INDArray ret = sameDiff.getArrForVarName(getVarName());
if(enforceExistence && ret == null) {
throw new IllegalStateException("No array exists for variable \"" + name() + "\"");
}
return ret;
}
/**
* Alias for the gradient variable - same as {@link #getGradient()}.
* The gradient variable is the variable that represents the derivative of the loss function with respect
* to the output of this variable. I.e., if this variable is X and loss function is L, then gradient() returns the
* variable representing dL/dX.
* Note that only floating point variables can have gradients.
*/
public SDVariable gradient() {
return getGradient();
}
/**
* The gradient variable is the variable that represents the derivative of the loss function with respect
* to the output of this variable. I.e., if this variable is X and loss function is L, then gradient() returns the
* variable representing dL/dX
* Note that only floating point variables can have gradients.
* Note also that a gradient may not yet be defined, and/or if no loss function variables have been set.
* You can set the loss function variables using {@link SameDiff#setLossVariables(String...)} and then create the
* gradient functions using {@link SameDiff#createGradFunction()}. Alternatively, the gradient function will be
* created automatically when training is performed.
*/
public SDVariable getGradient() {
Preconditions.checkState(dataType().isFPType(), "Cannot get gradient of %s datatype variable \"%s\": only floating" +
" point variables have gradients", dataType(), getVarName());
return sameDiff.getGradForVariable(getVarName());
}
/**
* Returns the shape of this variable
* @return Shape of the variable
*/
public long[] getShape() {
if (variableType == VariableType.PLACEHOLDER || sameDiff.isEagerMode() && shape != null) {
return shape;
} else if(variableType == VariableType.VARIABLE || variableType == VariableType.CONSTANT) {
if(getArr() != null)
return getArr().shape();
}
return null;
}
public void setShape(long... shape) {
this.shape = shape;
}
public long[] placeholderShape(){
if(variableType != VariableType.PLACEHOLDER){
throw new IllegalStateException("placeholderShape() can only be used for placeholder variables: variable \"" + getVarName()
+ " is a variable of type " + variableType);
}
return shape;
}
public DataType dataType() {
if(this.dataType == null) {
//Try to infer datatype instead of returning null
if(variableType != VariableType.ARRAY && getArr() != null) {
this.dataType = getArr().dataType();
} else {
this.dataType = DataType.UNKNOWN;
}
}
return this.dataType;
}
public LongShapeDescriptor getShapeDescriptor() {
return LongShapeDescriptor.fromShape(getShape(), this.dataType());
}
public SDVariable castTo(@NonNull DataType dataType){
return castTo(null, dataType);
}
public SDVariable castTo(String name, @NonNull DataType dataType){
return sameDiff.castTo(name, this, dataType);
}
/**
* Create a new SDVariable, the contents of which is copied from this current variable
* @return The new variable
*/
public SDVariable dup() {
return sameDiff.var(this);
}
/**
* Return a variable with equal shape to the input, but all elements set to the specified value
*
* @param value Value for returned variable
* @return new variable
*/
public SDVariable assign(Number value) {
return sameDiff.scalarSet(this, value.doubleValue());
}
/**
* Negate op - returns a new variable with the values of the current variable negated
* @return Negated variable
*/
public SDVariable neg(){
return sameDiff.math.neg(this);
}
/**
* Negate op - returns a new variable with the values of the current variable negated
* @param name Name of the new variable
* @return Negated variable
*/
public SDVariable neg(String name){
return sameDiff.math().neg(name, this);
}
/**
* See {@link #lt(String, double)}
*/
public SDVariable lt(double value){
return lt(null, value);
}
/**
* Less than operation: elementwise {@code this < value}
* Returns an array with the same shape/size as the input, with values 1 where condition is satisfied, or
* value 0 otherwise
*
* @param name Name of the output variable
* @param value value argument to use in operation
* @return Output SDVariable with values 0 (not satisfied) and 1 (where the condition is satisfied)
*/
public SDVariable lt(String name, double value){
return sameDiff.lt(name, this, value);
}
/**
* See {@link #lte(String, double)}
*/
public SDVariable lte(double value){
return lte(null, value);
}
/**
* Less than or equals operation: elementwise {@code this <= value}
* Returns an array with the same shape/size as the input, with values 1 where condition is satisfied, or
* value 0 otherwise
*
* @param name Name of the output variable
* @param value value argument to use in operation
* @return Output SDVariable with values 0 (not satisfied) and 1 (where the condition is satisfied)
*/
public SDVariable lte(String name, double value){
return sameDiff.lte(name, this, value);
}
/**
* See {@link #gt(String, double)}
*/
public SDVariable gt(double value){
return gt(null, value);
}
/**
* Greater than operation: elementwise {@code this > value}
* Returns an array with the same shape/size as the input, with values 1 where condition is satisfied, or
* value 0 otherwise
*
* @param name Name of the output variable
* @param value value argument to use in operation
* @return Output SDVariable with values 0 (not satisfied) and 1 (where the condition is satisfied)
*/
public SDVariable gt(String name, double value){
return sameDiff.gt(name, this, value);
}
/**
* See {@link #gte(String, double)}
*/
public SDVariable gte(double value){
return gte(null, value);
}
/**
* Greater than or equals operation: elementwise {@code this >= value}
* Returns an array with the same shape/size as the input, with values 1 where condition is satisfied, or
* value 0 otherwise
*
* @param name Name of the output variable
* @param value value argument to use in operation
* @return Output SDVariable with values 0 (not satisfied) and 1 (where the condition is satisfied)
*/
public SDVariable gte(String name, double value){
return sameDiff.gte(name, this, value);
}
/**
* See {@link #eq(String, double)}
*/
public SDVariable eq(double value){
return eq(null, value);
}
/**
* Equals operation: elementwise {@code this == value}
* Returns an array with the same shape/size as the input, with values 1 where condition is satisfied, or
* value 0 otherwise
*
* @param name Name of the output variable
* @param value value argument to use in operation
* @return Output SDVariable with values 0 (not satisfied) and 1 (where the condition is satisfied)
*/
public SDVariable eq(String name, double value){
return sameDiff.eq(name, this, value);
}
/**
* See {@link #neq(SDVariable)}
*/
public SDVariable neq(double value){
return neq(null, value);
}
/**
* Not equals operation: elementwise {@code this != value}
* Returns an array with the same shape/size as the input, with values 1 where condition is satisfied, or
* value 0 otherwise
*
* @param name Name of the output variable
* @param value value argument to use in operation
* @return Output SDVariable with values 0 (not satisfied) and 1 (where the condition is satisfied)
*/
public SDVariable neq(String name, double value){
return sameDiff.neq(name, this, value);
}
/**
* See {@link #lt(String, SDVariable)}
*/
public SDVariable lt(SDVariable other){
return lt(null, other);
}
/**
* Less than operation: elementwise {@code this < y}
* If x and y arrays have equal shape, the output shape is the same as the inputs.
* Supports broadcasting: if x and y have different shapes and are broadcastable, the output shape is broadcast.
* Returns an array with values 1 where condition is satisfied, or value 0 otherwise.
*
* @param name Name of the output variable
* @param other Variable to compare values against
* @return Output SDVariable with values 0 (not satisfied) and 1 (where the condition is satisfied)
*/
public SDVariable lt(String name, SDVariable other){
return sameDiff.lt(name, this, other);
}
/**
* See {@link #lte(String, SDVariable)}
*/
public SDVariable lte(SDVariable other){
return lte(null, other);
}
/**
* Less than or equal to operation: elementwise {@code this <= y}
* If x and y arrays have equal shape, the output shape is the same as the inputs.
* Supports broadcasting: if x and y have different shapes and are broadcastable, the output shape is broadcast.
* Returns an array with values 1 where condition is satisfied, or value 0 otherwise.
*
* @param name Name of the output variable
* @param other Variable to compare values against
* @return Output SDVariable with values 0 (not satisfied) and 1 (where the condition is satisfied)
*/
public SDVariable lte(String name, SDVariable other){
return sameDiff.lte(name, this, other);
}
/**
* See {@link #gt(String, SDVariable)}
*/
public SDVariable gt(SDVariable other){
return gt(null, other);
}
/**
* Greater than operation: elementwise {@code this > y}
* If x and y arrays have equal shape, the output shape is the same as the inputs.
* Supports broadcasting: if x and y have different shapes and are broadcastable, the output shape is broadcast.
* Returns an array with values 1 where condition is satisfied, or value 0 otherwise.
*
* @param name Name of the output variable
* @param other Variable to compare values against
* @return Output SDVariable with values 0 (not satisfied) and 1 (where the condition is satisfied)
*/
public SDVariable gt(String name, SDVariable other){
return sameDiff.gt(name, this, other);
}
/**
* See {@link #gte(String, SDVariable)}
*/
public SDVariable gte(SDVariable other){
return gte(null, other);
}
/**
* Greater than or equal to operation: elementwise {@code this >= y}
* If x and y arrays have equal shape, the output shape is the same as the inputs.
* Supports broadcasting: if x and y have different shapes and are broadcastable, the output shape is broadcast.
* Returns an array with values 1 where condition is satisfied, or value 0 otherwise.
*
* @param name Name of the output variable
* @param other Variable to compare values against
* @return Output SDVariable with values 0 (not satisfied) and 1 (where the condition is satisfied)
*/
public SDVariable gte(String name, SDVariable other){
return sameDiff.gte(name, this, other);
}
/**
* See {@link #eq(String, SDVariable)}
*/
public SDVariable eq(SDVariable other){
return eq(null, other);
}
/**
* Equal to operation: elementwise {@code this == y}
* If x and y arrays have equal shape, the output shape is the same as the inputs.
* Supports broadcasting: if x and y have different shapes and are broadcastable, the output shape is broadcast.
* Returns an array with values 1 where condition is satisfied, or value 0 otherwise.
*
* @param name Name of the output variable
* @param other Variable to compare values against
* @return Output SDVariable with values 0 (not satisfied) and 1 (where the condition is satisfied)
*/
public SDVariable eq(String name, SDVariable other){
return sameDiff.eq(name, this, other);
}
/**
* See {@link #neq(String, SDVariable)}
*/
public SDVariable neq(SDVariable other){
return neq(null, other);
}
/**
* Not equal to operation: elementwise {@code this != y}
* If x and y arrays have equal shape, the output shape is the same as the inputs.
* Supports broadcasting: if x and y have different shapes and are broadcastable, the output shape is broadcast.
* Returns an array with values 1 where condition is satisfied, or value 0 otherwise.
*
* @param name Name of the output variable
* @param other Variable to compare values against
* @return Output SDVariable with values 0 (not satisfied) and 1 (where the condition is satisfied)
*/
public SDVariable neq(String name, SDVariable other){
return sameDiff.neq(name, this, other);
}
/**
* See {@link #mmul(String, SDVariable)}
*/
public SDVariable mmul(SDVariable other){
return mmul(null, other);
}
/**
* Matrix multiplication: out = mmul(this,other)
*
* @param name Name of the output variable
* @param other Other variable to perform matrix multiplication with
* @return Output variable (result of mmul)
*/
public SDVariable mmul(String name, SDVariable other){
return sameDiff.mmul(name, this, other);
}
/**
* Matrix multiplication: out = mmul(this,other)
*
* @param name Name of the output variable
* @param other Other variable to perform matrix multiplication with
* @param mMulTranspose Matrix transpose configuration
* @return Output variable (result of mmul)
*/
public SDVariable mmul(String name, SDVariable other, @NonNull MMulTranspose mMulTranspose) {
return sameDiff.mmul(name, this, other, mMulTranspose.isTransposeA(), mMulTranspose.isTransposeB(), mMulTranspose.isTransposeResult());
}
/**
* See {@link #dot(String, SDVariable, int...)}
*/
public SDVariable dot(SDVariable other, int... dimensions){
return dot(null, other, dimensions);
}
/**
* Matrix dot product: out = dot(this,other, dimensions)
*
* @param name Name of the output variable
* @param other Other variable to perform matrix multiplication with
* @return Output variable (result of mmul)
*/
public SDVariable dot(String name, SDVariable other, int... dimensions){
return sameDiff.dot(name, this, other, dimensions);
}
/**
* See {@link #add(String, double)}
*/
public SDVariable add(double scalar) {
return add(null,scalar);
}
/**
* Scalar addition: {@code out = this + scalar}
* Output variable has the same shape as the input variable
*
* @param varName Output variable name
* @param scalar Scalar for operation
* @return Output variable
*/
public SDVariable add(String varName, double scalar) {
val function = sameDiff.math.add(this,scalar);
return sameDiff.updateVariableNameAndReference(function,varName);
}
/**
* See {@link #add(String, SDVariable)}
*/
public SDVariable add(SDVariable other) {
return add(null,other);
}
/**
* Addition operation: elementwise {@code this + x}
* If this and x variables have equal shape, the output shape is the same as the inputs.
* Supports broadcasting: if this and x have different shapes and are broadcastable, the output shape is broadcast.
*
* @param name Name of the output variable
* @param x Variable to perform operation with
* @return Output (result) SDVariable
*/
public SDVariable add(String name, SDVariable x) {
val result = sameDiff.math.add(this, x);
return sameDiff.updateVariableNameAndReference(result, name);
}
/**
* For Kotlin operator interop
* @see #add(String, SDVariable)
*/
public SDVariable plus(SDVariable other){
return add(other);
}
/**
* For Kotlin operator interop
* @see #add(String, double)
*/
public SDVariable plus(double other){
return add(other);
}
/**
* See {@link #sub(String, double)}
*/
public SDVariable sub(double scalar) {
return sub(null,scalar);
}
/**
* Scalar subtraction: {@code out = this - scalar}
* Output variable has the same shape as the input variable
*
* @param varName Output variable name
* @param scalar Scalar for operation
* @return Output variable
*/
public SDVariable sub(String varName, double scalar) {
val result = sameDiff.math.sub(this, scalar);
return sameDiff.updateVariableNameAndReference(result, varName);
}
/**
* See {@link #sub(String, SDVariable)}
*/
public SDVariable sub(SDVariable x) {
return sub(null,x);
}
/**
* Subtraction operation: elementwise {@code this - x}
* If this and x variables have equal shape, the output shape is the same as the inputs.
* Supports broadcasting: if this and x have different shapes and are broadcastable, the output shape is broadcast.
*
* @param name Name of the output variable
* @param x Variable to perform operation with
* @return Output (result) SDVariable
*/
public SDVariable sub(String name, SDVariable x) {
val result = sameDiff.math.sub(this,x);
return sameDiff.updateVariableNameAndReference(result,name);
}
/**
* For Kotlin operator interop
* @see #sub(String, SDVariable)
*/
public SDVariable minus(SDVariable other){
return sub(other);
}
/**
* For Kotlin operator interop
* @see #sub(String, double)
*/
public SDVariable minus(double other){
return sub(other);
}
/**
* See {@link #div(String,double)}
*/
public SDVariable div(double scalar) {
return div(null,scalar);
}
/**
* Scalar division: {@code out = this / scalar}
* Output variable has the same shape as the input variable
*
* @param varName Output variable name
* @param scalar Scalar for operation
* @return Output variable
*/
public SDVariable div(String varName, double scalar) {
val function = sameDiff.math.div(this,scalar);
return sameDiff.updateVariableNameAndReference(function,varName);
}
/**
* See {@link #div(String, SDVariable)}
*/
public SDVariable div(SDVariable x) {
return div(null,x);
}
/**
* Division operation: elementwise {@code this / x}
* If this and x variables have equal shape, the output shape is the same as the inputs.
* Supports broadcasting: if this and x have different shapes and are broadcastable, the output shape is broadcast.
*
* @param name Name of the output variable
* @param x Variable to perform operation with
* @return Output (result) SDVariable
*/
public SDVariable div(String name, SDVariable x) {
val result = sameDiff.math.div(this, x);
return sameDiff.updateVariableNameAndReference(result, name);
}
/**
* Floor division operation: elementwise {@code this // x}
* If this and x variables have equal shape, the output shape is the same as the inputs.
* Supports broadcasting: if this and x have different shapes and are broadcastable, the output shape is broadcast.
*
* @param name Name of the output variable
* @param x Variable to perform operation with
* @return Output (result) SDVariable
*/
public SDVariable fdiv(String name, SDVariable x) {
val result = sameDiff.math.floorDiv(this, x);
return sameDiff.updateVariableNameAndReference(result, name);
}
/**
* Modulo operation: elementwise {@code this / x}
* If this and x variables have equal shape, the output shape is the same as the inputs.
* Supports broadcasting: if this and x have different shapes and are broadcastable, the output shape is broadcast.
*
* @param name Name of the output variable
* @param x Variable to perform operation with
* @return Output (result) SDVariable
*/
public SDVariable mod(String name, SDVariable x) {
val result = sameDiff.math.mod(this, x);
return sameDiff.updateVariableNameAndReference(result, name);
}
/**
* See {@link #mul(String, double)}
*/
public SDVariable mul(double scalar) {
return mul(null,scalar);
}
/**
* Scalar multiplication: {@code out = this * scalar}
* Output variable has the same shape as the input variable
*
* @param varName Output variable name
* @param scalar Scalar for operation
* @return Output variable
*/
public SDVariable mul(String varName, double scalar) {
val function = sameDiff.math.mul(this, scalar);
return sameDiff.updateVariableNameAndReference(function,varName);
}
/**
* See {@link #mul(String, SDVariable)}
*/
public SDVariable mul(SDVariable x) {
return mul(null,x);
}
/**
* Multiplication operation: elementwise {@code this * x}
* If this and x variables have equal shape, the output shape is the same as the inputs.
* Supports broadcasting: if this and x have different shapes and are broadcastable, the output shape is broadcast.
*
* @param name Name of the output variable
* @param x Variable to perform operation with
* @return Output (result) SDVariable
*/
public SDVariable mul(String name, SDVariable x) {
val result = sameDiff.math.mul(this, x);
return sameDiff.updateVariableNameAndReference(result,name);
}
/**
* For Kotlin operator interop
* @see #mul(String, SDVariable)
*/
public SDVariable times(SDVariable other){
return mul(other);
}
/**
* For Kotlin operator interop
* @see #mul(String, double)
*/
public SDVariable times(double other){
return mul(other);
}
/**
* See {@link #pow(String, double)}
*/
public SDVariable pow(double scalar) {
return pow(null, scalar);
}
/**
* Scalar power operation: {@code out = this ^ scalar}
* Output variable has the same shape as the input variable
*
* @param varName Output variable name
* @param scalar Scalar for operation
* @return Output variable
*/
public SDVariable pow(String varName, double scalar) {
SDVariable ret = sameDiff.math.pow(this, scalar);
return sameDiff.updateVariableNameAndReference(ret, varName);
}
/**
* See {@link #rsub(String, double)}
*/
public SDVariable rsub(double scalar) {
return rsub(null,scalar);
}
/**
* Scalar reverse subtraction: {@code out = scalar - this}
* Output variable has the same shape as the input variable
*
* @param varName Output variable name
* @param scalar Scalar for operation
* @return Output variable
*/
public SDVariable rsub(String varName, double scalar) {
val function = sameDiff.math.rsub(this,scalar);
return sameDiff.updateVariableNameAndReference(function,varName);
}
/**
* See {@link #rsub(String, SDVariable)}
*/
public SDVariable rsub(SDVariable x) {
return rsub(null,x);
}
/**
* Reverse subtraction operation: elementwise {@code x - this}
* If this and x variables have equal shape, the output shape is the same as the inputs.
* Supports broadcasting: if this and x have different shapes and are broadcastable, the output shape is broadcast.
*
* @param name Name of the output variable
* @param x Variable to perform operation with
* @return Output (result) SDVariable
*/
public SDVariable rsub(String name, SDVariable x) {
val result = sameDiff.math.rsub(this,x);
return sameDiff.updateVariableNameAndReference(result,name);
}
/**
* See {@link #rdiv(String, double)}
*/
public SDVariable rdiv(double scalar) {
return rdiv(null,scalar);
}
/**
* Scalar reverse division: {@code out = scalar / this}
* Output variable has the same shape as the input variable
*
* @param varName Output variable name
* @param scalar Scalar for operation
* @return Output variable
*/
public SDVariable rdiv(String varName, double scalar) {
val function = sameDiff.math.rdiv(this, scalar);
return sameDiff.updateVariableNameAndReference(function, varName);
}
/**
* See {@link #rdiv(String, SDVariable)}
*/
public SDVariable rdiv(SDVariable sameDiffVariable) {
return rdiv(null,sameDiffVariable);
}
/**
* Reverse division operation: elementwise {@code x / this}
* If this and x variables have equal shape, the output shape is the same as the inputs.
* Supports broadcasting: if this and x have different shapes and are broadcastable, the output shape is broadcast.
*
* @param name Name of the output variable
* @param x Variable to perform operation with
* @return Output (result) SDVariable
*/
public SDVariable rdiv(String name, SDVariable x) {
val result = sameDiff.math.rdiv(this,x);
return sameDiff.updateVariableNameAndReference(result,name);
}
/**
* See {@link #squaredDifference(String, SDVariable)}
*/
public SDVariable squaredDifference(SDVariable x) {
return squaredDifference(null,x);
}
/**
* Squared difference operation: {@code (this - x)^2}
* @param x Other input variable
* @return squared difference between variables
*/
public SDVariable squaredDifference(String name, SDVariable x) {
val result = sameDiff.math().squaredDifference(this, x);
return sameDiff.updateVariableNameAndReference(result, name);
}
/**
* See {@link #sum(String, boolean, int...)}
*/
public SDVariable sum(int... dimensions){
return sum(null, dimensions);
}
/**
* See {@link #sum(String, boolean, int...)}
*/
public SDVariable sum(boolean keepDims, int... dimensions){
return sum(null, keepDims, dimensions);
}
/**
* See {@link #sum(String, boolean, int...)}
*/
public SDVariable sum(String name, int... dimensions){
return sum(name, false, dimensions);
}
/**
* Sum array reduction operation, optionally along specified dimensions.
* Note that if keepDims = true, the output variable has the same rank as the input variable,
* with the reduced dimensions having size 1. This can be useful for later broadcast operations (such as subtracting
* the mean along a dimension).
* Example: if input has shape [a,b,c] and dimensions=[1] then output has shape:
* keepDims = true: [a,1,c]
* keepDims = false: [a,c]
*
* @param name Output variable name
* @param keepDims If true: keep the dimensions that are reduced on (as length 1). False: remove the reduction dimensions
* @param dimensions Dimensions to reduce over. If dimensions are not specified, full array reduction is performed
* @return Output variable: reduced array of rank (input rank - num dimensions) if keepDims = false, or
* of rank (input rank) if keepdims = true
*/
public SDVariable sum(String name, boolean keepDims, int... dimensions){
return sameDiff.sum(name, this, keepDims, dimensions);
}
/**
* See {@link #mean(String, boolean, int...)}
*/
public SDVariable mean(boolean keepDims, int... dimensions){
return mean(null, keepDims, dimensions);
}
/**
* See {@link #mean(String, boolean, int...)}
*/
public SDVariable mean(String name, int... dimensions){
return mean(name, false, dimensions);
}
/**
* See {@link #mean(String, boolean, int...)}
*/
public SDVariable mean(int... dimensions){
return mean(null, false, dimensions);
}
/**
* Mean (average) array reduction operation, optionally along specified dimensions
* Note that if keepDims = true, the output variable has the same rank as the input variable,
* with the reduced dimensions having size 1. This can be useful for later broadcast operations (such as subtracting
* the mean along a dimension).
* Example: if input has shape [a,b,c] and dimensions=[1] then output has shape:
* keepDims = true: [a,1,c]
* keepDims = false: [a,c]
*
* @param name Output variable name
* @param keepDims If true: keep the dimensions that are reduced on (as size 1). False: remove the reduction dimensions
* @param dimensions Dimensions to reduce over. If dimensions are not specified, full array reduction is performed
* @return Reduced array of rank (input rank - num dimensions)
*/
public SDVariable mean(String name, boolean keepDims, int... dimensions){
return sameDiff.mean(name, this, keepDims, dimensions);
}
/**
* See {@link #std(String, boolean, boolean, int...)}
*/
public SDVariable std(boolean biasCorrected, int... dimensions){
return std(null, biasCorrected, dimensions);
}
/**
* See {@link #std(String, boolean, boolean, int...)}
*/
public SDVariable std(String name, boolean biasCorrected, int... dimensions){
return sameDiff.standardDeviation(name, this, biasCorrected, dimensions);
}
/**
* Stardard deviation array reduction operation, optionally along specified dimensions
* Note that if keepDims = true, the output variable has the same rank as the input variable,
* with the reduced dimensions having size 1. This can be useful for later broadcast operations (such as subtracting
* the mean along a dimension).
* Example: if input has shape [a,b,c] and dimensions=[1] then output has shape:
* keepDims = true: [a,1,c]
* keepDims = false: [a,c]
*
* @param biasCorrected If true: divide by (N-1) (i.e., sample stdev). If false: divide by N (population stdev)
* @param keepDims If true: keep the dimensions that are reduced on (as size 1). False: remove the reduction dimensions
* @param dimensions Dimensions to reduce over. If dimensions are not specified, full array reduction is performed
* @return Output variable: reduced array of rank (input rank - num dimensions)
*/
public SDVariable std(String name, boolean biasCorrected, boolean keepDims, int... dimensions){
return sameDiff.standardDeviation(name, this, biasCorrected, keepDims, dimensions);
}
/**
* See {@link #prod(String, boolean, int...)}
*/
public SDVariable prod(int... dimensions){
return prod(null, dimensions);
}
/**
* See {@link #prod(String, boolean, int...)}
*/
public SDVariable prod(boolean keepDims, int... dimensions){
return prod(null, keepDims, dimensions);
}
/**
* See {@link #prod(String, boolean, int...)}
*/
public SDVariable prod(String name, int... dimensions){
return sameDiff.prod(name, this, dimensions);
}
/**
* Product array reduction operation, optionally along specified dimensions
* Note that if keepDims = true, the output variable has the same rank as the input variable,
* with the reduced dimensions having size 1. This can be useful for later broadcast operations (such as subtracting
* the mean along a dimension).
* Example: if input has shape [a,b,c] and dimensions=[1] then output has shape:
* keepDims = true: [a,1,c]
* keepDims = false: [a,c]
*
* @param name Output variable name
* @param keepDims If true: keep the dimensions that are reduced on (as length 1). False: remove the reduction dimensions
* @param dimensions Dimensions to reduce over. If dimensions are not specified, full array reduction is performed
* @return Output variable: reduced array of rank (input rank - num dimensions)
*/
public SDVariable prod(String name, boolean keepDims, int... dimensions){
return sameDiff.prod(name, this, keepDims, dimensions);
}
/**
* See {@link #min(String, boolean, int...)}
*/
public SDVariable min(int... dimensions){
return min(null, dimensions);
}
/**
* See {@link #min(String, boolean, int...)}
*/
public SDVariable min(boolean keepDims, int... dimensions){
return min(null, keepDims, dimensions);
}
/**
* See {@link #min(String, boolean, int...)}
*/
public SDVariable min(String name, int... dimensions){
return min(name, false, dimensions);
}
/**
* Minimum array reduction operation, optionally along specified dimensions. out = min(in)
* Note that if keepDims = true, the output variable has the same rank as the input variable,
* with the reduced dimensions having size 1. This can be useful for later broadcast operations (such as subtracting
* the mean along a dimension).
* Example: if input has shape [a,b,c] and dimensions=[1] then output has shape:
* keepDims = true: [a,1,c]
* keepDims = false: [a,c]
*
* @param name Output variable name
* @param keepDims If true: keep the dimensions that are reduced on (as size 1). False: remove the reduction dimensions
* @param dimensions Dimensions to reduce over. If dimensions are not specified, full array reduction is performed
* @return Reduced array of rank (input rank - num dimensions)
*/
public SDVariable min(String name, boolean keepDims, int... dimensions){
return sameDiff.min(name, this, keepDims, dimensions);
}
/**
* See {@link #max(String, boolean, int...)}
*/
public SDVariable max(int... dimensions) {
return max(null, dimensions);
}
/**
* See {@link #max(String, boolean, int...)}
*/
public SDVariable max(String name, int... dimensions) {
return max(name, false, dimensions);
}
/**
* See {@link #max(String, boolean, int...)}
*/
public SDVariable max(boolean keepDims, int... dimensions) {
return max(null, keepDims, dimensions);
}
/**
* Maximum array reduction operation, optionally along specified dimensions
* Note that if keepDims = true, the output variable has the same rank as the input variable,
* with the reduced dimensions having size 1. This can be useful for later broadcast operations (such as subtracting
* the mean along a dimension).
* Example: if input has shape [a,b,c] and dimensions=[1] then output has shape:
* keepDims = true: [a,1,c]
* keepDims = false: [a,c]
*
* @param name Output variable name
* @param keepDims If true: keep the dimensions that are reduced on (as size 1). False: remove the reduction dimensions
* @param dimensions Dimensions to reduce over. If dimensions are not specified, full array reduction is performed
* @return Reduced array of rank (input rank - num dimensions)
*/
public SDVariable max(String name, boolean keepDims, int... dimensions) {
return sameDiff.max(name, this, keepDims, dimensions);
}
/**
* See {@link #norm1(String, boolean, int...)}
*/
public SDVariable norm1(int... dimensions){
return norm1(null, dimensions);
}
/**
* See {@link #norm1(String, boolean, int...)}
*/
public SDVariable norm1(boolean keepDims, int... dimensions){
return norm1(null, keepDims, dimensions);
}
/**
* See {@link #norm1(String, boolean, int...)}
*/
public SDVariable norm1(String name, int... dimensions){
return norm1(name, false, dimensions);
}
/**
* Norm1 (L1 norm) reduction operation: The output contains the L1 norm for each tensor/subset along the specified dimensions:
* {@code out = sum_i abs(x[i])}
* Note that if keepDims = true, the output variable has the same rank as the input variable,
* with the reduced dimensions having size 1. This can be useful for later broadcast operations (such as subtracting
* the mean along a dimension).
* Example: if input has shape [a,b,c] and dimensions=[1] then output has shape:
* keepDims = true: [a,1,c]
* keepDims = false: [a,c]
*
* @param name Output variable name
* @param keepDims If true: keep the dimensions that are reduced on (as size 1). False: remove the reduction dimensions
* @param dimensions dimensions to reduce over
* @return Output variable
*/
public SDVariable norm1(String name, boolean keepDims, int... dimensions) {
return sameDiff.norm1(name, this, keepDims, dimensions);
}
/**
* See {@link #norm2(String, boolean, int...)}
*/
public SDVariable norm2(int... dimensions){
return norm2(null, dimensions);
}
/**
* See {@link #norm2(String, boolean, int...)}
*/
public SDVariable norm2(boolean keepDims, int... dimensions){
return norm2(null, keepDims, dimensions);
}
/**
* See {@link #norm2(String, boolean, int...)}
*/
public SDVariable norm2(String name, int... dimensions){
return norm2(name, false, dimensions);
}
/**
* Norm2 (L2 norm) reduction operation: The output contains the L2 norm for each tensor/subset along the specified dimensions:
* {@code out = sqrt(sum_i x[i]^2)}
* Note that if keepDims = true, the output variable has the same rank as the input variable,
* with the reduced dimensions having size 1. This can be useful for later broadcast operations (such as subtracting
* the mean along a dimension).
* Example: if input has shape [a,b,c] and dimensions=[1] then output has shape:
* keepDims = true: [a,1,c]
* keepDims = false: [a,c]
*
* @param name Output variable name
* @param keepDims If true: keep the dimensions that are reduced on (as size 1). False: remove the reduction dimensions
* @param dimensions dimensions to reduce over
* @return Output variable
*/
public SDVariable norm2(String name, boolean keepDims, int... dimensions){
return sameDiff.norm2(name, this, keepDims, dimensions);
}
/**
* See {@link #normmax(String, boolean, int...)}
*/
public SDVariable normmax(int... dimensions){
return normmax(null, dimensions);
}
/**
* See {@link #normmax(String, boolean, int...)}
*/
public SDVariable normmax(boolean keepDims, int... dimensions){
return normmax(null, keepDims, dimensions);
}
/**
* See {@link #normmax(String, boolean, int...)}
*/
public SDVariable normmax(String name, int... dimensions){
return normmax(name, false, dimensions);
}
/**
* Max norm (infinity norm) reduction operation: The output contains the max norm for each tensor/subset along the
* specified dimensions:
* {@code out = max(abs(x[i]))}
* Note that if keepDims = true, the output variable has the same rank as the input variable,
* with the reduced dimensions having size 1. This can be useful for later broadcast operations (such as subtracting
* the mean along a dimension).
* Example: if input has shape [a,b,c] and dimensions=[1] then output has shape:
* keepDims = true: [a,1,c]
* keepDims = false: [a,c]
*
* @param name Output variable name
* @param keepDims If true: keep the dimensions that are reduced on (as size 1). False: remove the reduction dimensions
* @param dimensions dimensions to reduce over
* @return Output variable
*/
public SDVariable normmax(String name, boolean keepDims, int... dimensions){
return sameDiff.normmax(name, this, keepDims, dimensions);
}
/**
* See {@link #argmax(String, boolean, int...)}
*/
public SDVariable argmax(int... dimensions){
return argmax(null, dimensions);
}
/**
* See {@link #argmax(String, boolean, int...)}
*/
public SDVariable argmax(String name, int... dimensions){
return sameDiff.argmax(name, this, dimensions);
}
/**
* Argmax array reduction operation, optionally along specified dimensions.
* Output values are the index of the maximum value of each slice along the specified dimension.
*
* Note that if keepDims = true, the output variable has the same rank as the input variable,
* with the reduced dimensions having size 1. This can be useful for later broadcast operations (such as subtracting
* the mean along a dimension).
* Example: if input has shape [a,b,c] and dimensions=[1] then output has shape:
* keepDims = true: [a,1,c]
* keepDims = false: [a,c]
*
* @param name Name of the output variable
* @param keepDims If true: keep the dimensions that are reduced on (as size 1). False: remove the reduction dimensions
* @param dimensions Dimensions to reduce over. If dimensions are not specified, full array reduction is performed
* @return Output variable: reduced array of rank (input rank - num dimensions) if keepDims = false, or
* of rank (input rank) if keepdims = true
*/
public SDVariable argmax(String name, boolean keepDims, int... dimensions) {
return sameDiff.argmax(name, this, keepDims, dimensions);
}
/**
* See {@link #argmin(String, boolean, int...)}
*/
public SDVariable argmin(int... dimensions){
return argmin(null, dimensions);
}
/**
* See {@link #argmin(String, boolean, int...)}
*/
public SDVariable argmin(String name, int... dimensions){
return sameDiff.argmin(name, this, dimensions);
}
/**
* Argmin array reduction operation, optionally along specified dimensions.
* Output values are the index of the minimum value of each slice along the specified dimension.
*
* Note that if keepDims = true, the output variable has the same rank as the input variable,
* with the reduced dimensions having size 1. This can be useful for later broadcast operations (such as subtracting
* the mean along a dimension).
* Example: if input has shape [a,b,c] and dimensions=[1] then output has shape:
* keepDims = true: [a,1,c]
* keepDims = false: [a,c]
*
* @param name Name of the output variable
* @param keepDims If true: keep the dimensions that are reduced on (as length 1). False: remove the reduction dimensions
* @param dimensions Dimensions to reduce over. If dimensions are not specified, full array reduction is performed
* @return Output variable: reduced array of rank (input rank - num dimensions) if keepDims = false, or
* of rank (input rank) if keepdims = true
*/
public SDVariable argmin(String name, boolean keepDims, int... dimensions) {
return sameDiff.argmax(name, this, keepDims, dimensions);
}
/**
* Return the total number of elements in this array
* @return
*/
public SDVariable length() {
return sameDiff.prod(shape());
}
/**
* Get the shape of the array as a dynamic SDVariable
* @return Shape SDVariable
*/
public SDVariable shape(){
return sameDiff.shape(this);
}
/**
* Get the rank of this variable as a dynamic SDVariable
* @return Rank SDVariable
*/
public SDVariable rank(){
return sameDiff.rank(this);
}
/**
* Reshape the current variable to the specified (dynamic) shape. The output variable will have the same values as the
* input, but with the specified shape.
* Note that prod(shape) must match length(input) == prod(input.shape)
*
* @param newShape New shape for variable
* @return Output variable
*/
public SDVariable reshape(SDVariable newShape) {
return sameDiff.reshape(this, newShape);
}
/**
* Reshape the current variable to the specified (dynamic) shape. The output variable will have the same values as the
* input, but with the specified shape.
* Note that prod(shape) must match length(input) == prod(input.shape)
*
* @param newShape New shape for variable
* @return Output variable
*/
public SDVariable reshape(String name,SDVariable newShape) {
return sameDiff.reshape(name,this, newShape);
}
/**
* Reshape the current variable to the specified shape. The output variable will have the same values as the
* input, but with the specified shape.
* Note that prod(shape) must match length(input) == prod(input.shape)
*
* @param newShape New shape for variable
* @return Output variable
*/
public SDVariable reshape(int... newShape){
return sameDiff.reshape(this, ArrayUtil.toLongArray(newShape));
}
/**
* Reshape the current variable to the specified shape. The output variable will have the same values as the
* input, but with the specified shape.
* Note that prod(shape) must match length(input) == prod(input.shape)
*
* @param newShape New shape for variable
* @return Output variable
*/
public SDVariable reshape(long... newShape){
return sameDiff.reshape(this, newShape);
}
/**
* Permute the dimensions of the current variable according to the specified permutation indices.
* Example: if the current variable has shape [a,b,c] and dimensions = [2,0,1] the output has shape [c,a,b]
*
* @param dimensions The new dimension order
* @return Output variable (permuted input)
*/
public SDVariable permute(int... dimensions){
return sameDiff.permute(this, dimensions);
}
public SDVariable permute(SDVariable dimensions){
return sameDiff.permute( this, dimensions);
}
/**
* Associate the specified array with this variable
* @param array Array to associate with this variable
* @return This variable
*/
public SDVariable setArray(INDArray array){
sameDiff.associateArrayWithVariable(array, this);
return this;
}
/**
* Evaluate the result of this variable
* @return
*/
public INDArray eval() {
Map m = sameDiff.output((Map)null, name());
return m.get(name());
}
/**
* Evaluate the result of this variable
* @return
*/
public INDArray eval(Map placeholders) {
Map m = sameDiff.output(placeholders, name());
return m.get(name());
}
@Override
public String toString() {
return "SDVariable(name=\"" + varName + "\",variableType=" + variableType + ",dtype=" + dataType +
(variableType == VariableType.PLACEHOLDER && shape != null ? ",shape=" + Arrays.toString(shape): "") + ")";
}
/**
* Add a control dependency for this variable on the specified variable.
* Control dependencies can be used to enforce the execution order.
* For example, if a control dependency X->Y exists, then Y will only be executed after X is executed - even
* if Y wouldn't normally depend on the result/values of X.
*
* @param controlDependency Control dependency to add for this variable
*/
public void addControlDependency(SDVariable controlDependency){
Variable vThis = sameDiff.getVariables().get(getVarName());
Variable vCD = sameDiff.getVariables().get(controlDependency.name());
//If possible: add control dependency on ops
if(vThis.getOutputOfOp() != null && vCD.getOutputOfOp() != null ){
//Op -> Op case
SameDiffOp oThis = sameDiff.getOps().get(vThis.getOutputOfOp());
SameDiffOp oCD = sameDiff.getOps().get(vCD.getOutputOfOp());
if(oThis.getControlDeps() == null)
oThis.setControlDeps(new ArrayList<>());
if(!oThis.getControlDeps().contains(oCD.getName()))
oThis.getControlDeps().add(oCD.getName());
if(oCD.getControlDepFor() == null)
oCD.setControlDepFor(new ArrayList<>());
if(!oCD.getControlDepFor().contains(oThis.getName()))
oCD.getControlDepFor().add(oThis.getName());
} else {
if(vThis.getOutputOfOp() != null){
//const/ph -> op case
SameDiffOp oThis = sameDiff.getOps().get(vThis.getOutputOfOp());
if(oThis.getVarControlDeps() == null)
oThis.setVarControlDeps(new ArrayList<>());
if(!oThis.getVarControlDeps().contains(vCD.getName()))
oThis.getVarControlDeps().add(vCD.getName());
if(vCD.getControlDepsForOp() == null)
vCD.setControlDepsForOp(new ArrayList<>());
if(!vCD.getControlDepsForOp().contains(oThis.getName()))
vCD.getControlDepsForOp().add(oThis.getName());
} else {
//const/ph -> const/ph case
if(vThis.getControlDeps() == null)
vThis.setControlDeps(new ArrayList<>());
if(!vThis.getControlDeps().contains(vCD.getName()))
vThis.getControlDeps().add(vCD.getName());
if(vCD.getControlDepsForVar() == null)
vCD.setControlDepsForVar(new ArrayList<>());
if(!vCD.getControlDepsForVar().contains(vThis.getName()))
vCD.getControlDepsForVar().add(vThis.getName());
}
}
}
/**
* Get a variable with content equal to a specified sub-array of this variable.
* Can be used (for example) to get rows, columns, sub-matrices, etc.
* @param indices Indices to get
* @return Sub-array variable
*/
public SDVariable getView(SDIndex... indices) {
//copy because we can mutate this internally
SDVariable[] indicesVars = new SDVariable[indices.length];
for(int i = 0; i < indices.length; i++) {
//convert indices to SDVariable based indices
switch(indices[i].getIndexType()) {
case INTERVAL:
indicesVars[i] = CreateView.createInterval(sameDiff,indices[i].getIntervalBegin(),indices[i].getIntervalEnd(),indices[i].getIntervalStrides(),indices[i].isInclusive() ? 1 : 0);
break;
case POINT:
indicesVars[i] = CreateView.createPoint(sameDiff,indices[i].getPointIndex());
break;
case POINT_INPUT:
indicesVars[i] = CreateView.createPoint(sameDiff,indices[i].getPointVar());
break;
case INTERVAL_INPUT:
indicesVars[i] = CreateView.createInterval(sameDiff,indices[i].getIntervalInputBegin(),indices[i].getIntervalInputEnd(),indices[i].getIntervalStrideInput(), indices[i].getInclusiveInput());
break;
case ALL:
indicesVars[i] = CreateView.createAll(sameDiff);
break;
default:
throw new IllegalArgumentException("Illegal type " + indices[i].getIndexType());
}
}
return sameDiff.createView(this,indicesVars);
}
/**
* Get a variable with content equal to a specified sub-array of this variable.
* Can be used (for example) to get rows, columns, sub-matrices, etc.
* @param indices Indices to get
* @return Sub-array variable
*/
public SDVariable get(SDIndex... indices) {
int ndims = indices.length;
boolean variableIndices = false;
//copy because we can mutate this internally
SDIndex[] inputIndices = Arrays.copyOf(indices,indices.length);
indices = inputIndices;
for(int i = 0; i < indices.length; i++) {
if(indices[i].getIndexType() == SDIndex.IndexType.POINT_INPUT || indices[i].getIndexType() == SDIndex.IndexType.INTERVAL_INPUT) {
variableIndices = true;
}
//convert indices to SDVariable based indices
if(variableIndices && (indices[i].getIndexType() == SDIndex.IndexType.INTERVAL || indices[i].getIndexType() == SDIndex.IndexType.POINT)) {
switch(indices[i].getIndexType()) {
case INTERVAL:
indices[i] = SDIndex.interval(sameDiff.constant(indices[i].getIntervalBegin()),sameDiff.constant(indices[i].getIntervalEnd()),sameDiff.constant(indices[i].getIntervalEnd()));
break;
case POINT:
indices[i] = SDIndex.point(sameDiff.constant(indices[i].getPointIndex()),indices[i].isPointKeepDim());
break;
}
}
}
long[] begin = new long[ndims];
long[] end = new long[ndims];
long[] strides = new long[ndims];
int[] begin_mask_arr = new int[ndims];
int[] end_mask_arr = new int[ndims];
int[] shrink_axis_mask_arr = new int[ndims];
SDVariable beginVar = null;
SDVariable endVar = null;
SDVariable stridesVar = null;
for (int i = 0; i < ndims; i++) {
strides[i] = 1;
SDIndex index = indices[i];
SDIndex.IndexType indexType = index.getIndexType();
if (indexType == SDIndex.IndexType.ALL) {
begin_mask_arr[i] = 1;
end_mask_arr[i] = 1;
} else if (indexType == SDIndex.IndexType.POINT || indexType == SDIndex.IndexType.POINT_INPUT) {
if(indexType == SDIndex.IndexType.POINT) {
long pointIndex = index.getPointIndex();
begin[i] = pointIndex;
end[i] = pointIndex + 1;
} else if(indexType == SDIndex.IndexType.POINT_INPUT) {
if(beginVar == null && endVar == null) {
beginVar = index.getPointVar();
endVar = index.getPointVar().add(1.0);
} else {
beginVar = sameDiff.concat(0,beginVar,index.getPointVar());
endVar = sameDiff.concat(0,endVar,index.getPointVar().add(1.0));
}
}
if(!index.isPointKeepDim()) {
shrink_axis_mask_arr[i] = 1;
}
} else if (indexType == SDIndex.IndexType.INTERVAL || indexType == SDIndex.IndexType.INTERVAL_INPUT) {
if (index.getIntervalBegin() == null && indexType != SDIndex.IndexType.INTERVAL_INPUT) {
begin_mask_arr[i] = 1;
} else if(indexType == SDIndex.IndexType.INTERVAL_INPUT) {
if(beginVar == null) {
beginVar = index.getIntervalInputBegin();
} else {
beginVar = sameDiff.concat(0,beginVar,index.getIntervalInputBegin());
}
} else {
begin[i] = index.getIntervalBegin();
}
if (index.getIntervalEnd() == null && indexType != SDIndex.IndexType.INTERVAL_INPUT) {
end_mask_arr[i] = 1;
} else if(indexType == SDIndex.IndexType.INTERVAL_INPUT) {
if(endVar == null) {
endVar = index.getIntervalInputEnd();
} else {
endVar = sameDiff.concat(0,endVar,index.getIntervalInputEnd());
}
} else {
end[i] = index.getIntervalEnd();
}
if (index.getIntervalStrides() == null) {
strides[i] = 1;
if(stridesVar != null) {
stridesVar = sameDiff.concat(0,stridesVar,sameDiff.constant(1).reshape(1));
} else {
stridesVar = sameDiff.constant(1).reshape(1);
}
} else {
strides[i] = index.getIntervalStrides();
if(stridesVar != null) {
stridesVar = sameDiff.concat(0,stridesVar,index.getIntervalStrideInput());
} else {
stridesVar = index.getIntervalStrideInput();
}
}
}
}
// convert binary int[] to int
int begin_mask = binArrToInt(begin_mask_arr);
int end_mask = binArrToInt(end_mask_arr);
int shrink_axis = binArrToInt(shrink_axis_mask_arr);
if(variableIndices) {
if(stridesVar == null) {
stridesVar = sameDiff.onesLike(beginVar);
}
return this.sameDiff.stridedSlice(this, beginVar, endVar, stridesVar,
begin_mask, end_mask, 0, 0, shrink_axis);
} else {
return this.sameDiff.stridedSlice(this, begin, end, strides,
begin_mask, end_mask, 0, 0, shrink_axis);
}
}
public static SDVariable sliceEnd(SDVariable input,SDVariable sliceIndexInput) {
SameDiff sameDiff = input.getSameDiff();
SDVariable range = sameDiff.range(sameDiff.constant(0), input.rank(), sameDiff.constant(1), DataType.INT64);
//0 1 1
SDVariable mask = range.gt(0.0).castTo(DataType.INT64);
SDVariable sliceMask = range.eq(0).castTo(DataType.INT64);
SDVariable sliceIndex = sliceMask.mul(sliceIndexInput);
SDVariable outputShape = input.shape().mul(mask).add(sliceIndex);
return outputShape;
}
/**
* Get a variable with content equal to a specified sub-array of this variable.
* Can be used (for example) to get rows, columns, sub-matrices, etc.
*
* This will loop over the indices (think of it as a list) and concatenate
* each slice of the input array to the final result.
*
* Expected input for indices would be a vector with indices such as 0,1,2,3,4.
* For each element in the index we then concatenate the result to the previous iteration.
*
* Note that this is slow and should only be used in very specific circumstances.
* Otherwise {@link org.nd4j.linalg.api.ops.impl.shape.StridedSlice} will be more performant
* for creating views. Many times {@link org.nd4j.linalg.api.ops.impl.shape.StridedSlice} avoids
* this slower approach by directly calculating the strides of a view.
*
* @param indices Indices to get
* @return Sub-array variable
*/
public SDVariable get(SDVariable indices) {
SDVariable initialSize = sameDiff.zerosLike(shape()).castTo(DataType.INT64);
//pull from the first slice as the starting point and concatenate each result together
SDVariable startResult = sameDiff.slice(this, initialSize.castTo(DataType.INT64), sliceEnd(this,
sameDiff.onesLike(shape()).castTo(DataType.INT64)));
//start at 1 because we start with the initial output (basically the item at the first element in the indices)
SDVariable currIteration = sameDiff.var(Nd4j.ones(1).castTo(DataType.INT32));
//this condition is normally used when you want to toss in an extra condition to terminate early
SDVariable cond = sameDiff.constant("curr_cond",true);
//the total length of the indices to loop till
SDVariable indicesLength = indices.length();
//sub graph that uses invoke
SameDiff loop = createLoopConcat(this,indices);
//collect slices along the first dimension concatenating the result along the way
return this.sameDiff.loopWithConditions(ControlFlow.LoopParams.builder()
.functionBody(loop)
.loopVars(new SDVariable[] {
currIteration,
indicesLength,
cond,
startResult,
this,
indices
}).functionBodyInputs(new String[] {
//note here all inputs are the same as the outputs, and we return the original
//input concatenated with the starting input (the first slice at index 0)
//and then loop over each index in the list till we get the specific result
"index",
"max",
"cond",
"input",
"pullFrom",
"indices"
})
.functionBodyOutputs(new String[]{
"index",
"max",
"cond",
"output",
"pullFrom",
"indices"})
.functionName("slices")
.loopName("outputs")
//note the ordering here is important. Output is the accumulated output of each iteration appending
//a result to the previous iteration. We start with the initial input and add more overtime.
.build())[3];
}
/**
* Get a variable with content equal to a specified sub-array of this variable.
* Can be used (for example) to get rows, columns, sub-matrices, etc.
*
* This will loop over the indices (think of it as a list) and add each slice
* specified by the indices from the source to the new array.
*
* The end result will be this variable but with the new updated results.
*
* Note that this is slow and should only be used in very specific circumstances.
* Otherwise {@link org.nd4j.linalg.api.ops.impl.shape.StridedSlice} will be more performant
* for creating views. Many times {@link org.nd4j.linalg.api.ops.impl.shape.StridedSlice} avoids
* this slower approach by directly calculating the strides of a view.
*
* @param indices Indices to get
* @param toPut the source array to pull results from to put in to this array
* @param putIndices the equivalent indices for the other array
* @return the updated array with the elements from the toPut array put in to this new array
*/
public SDVariable put(SDVariable indices,SDVariable toPut,SDVariable putIndices) {
//start at 1 because we start with the initial output (basically the item at the first element in the indices)
SDVariable currIteration = sameDiff.var(Nd4j.zeros(1).castTo(DataType.INT32));
//this condition is normally used when you want to toss in an extra condition to terminate early
SDVariable cond = sameDiff.constant(true);
//the total length of the indices to loop till
SDVariable indicesLength = indices.length();
//sub graph that uses invoke
SameDiff loop = createLoopPut(this,indices);
loop.setEnableCache(false);
//collect slices along the first dimension concatenating the result along the way
return this.sameDiff.loopWithConditions(ControlFlow.LoopParams.builder()
.functionBody(loop)
.loopVars(new SDVariable[] {
currIteration,
indicesLength,
cond,
this,
toPut,
indices,
putIndices
}).functionBodyInputs(new String[] {
//note here all inputs are the same as the outputs, and we return the original
//the default 3 values (current iteration, max index to loop to and optional condition)
//index,max,cond,assignTo,putIn,indices,indicesPut
"index",
"max",
"cond",
"assignTo",
"toPut",
"indices",
"indicesPut"
})
.functionBodyOutputs(new String[]{
"index",
"max",
"cond",
"assignOutput",
"toPut",
"indices",
"indicesPut"})
.functionName("sliceputs")
.loopName("outputs")
//note the ordering here is important. Output is the original array where we assigned values.
.build())[3];
}
/**
* Create a graph that takes in the indices as a placeholder, loops over each element in the index vector
* and appends the slice to the end result. This graph is equivalent to something like:
* INDArray input = ....;
* INDArray indices = ...;
* INDArray result = input.get(NDArrayIndex.point(indices.getInt(0));
* for(int i = i; i < maxIndex && customInputResult; i++) {
* result = Nd4j.concat(0,input.get(NDArrayIndex.point(i)));
* }
* return result
*
* Note this is similar to {@link INDArray#get(INDArray)}
*
* @param relative the expected target input variable. We use this to pull expected
* return data type for the result
* @param indices the indices to get
* @return the graph for dynamically creating a result graph
*/
public static SameDiff createLoopPut(SDVariable relative,SDVariable indices) {
//standard loop body for loopWithConditions
SameDiff loop = SameDiff.create();
//curr index
SDVariable index = loop.placeHolder("index",DataType.INT32);
//loop until
SDVariable maxIndex = loop.placeHolder("max",DataType.INT32);
//constant condition of true for custom, just loop till max iterations hit
SDVariable currCondition = loop.placeHolder("cond",DataType.BOOL);
//the actual variable to pull from
SDVariable assignTo = loop.placeHolder("assignTo",relative.dataType());
SDVariable toPut = loop.placeHolder("toPut",relative.dataType());
//the indices to loop over (the input variable
SDVariable indicesLoop = loop.placeHolder("indices",indices.dataType());
//standardize indices to length 1
indicesLoop = indicesLoop.reshape("indicesReshape",indicesLoop.length());
SDVariable indicesPut = loop.placeHolder("indicesPut",indices.dataType());
indicesPut = indicesPut.reshape("indicesPutReshape",indicesPut.length());
//the current index to retrieve
SDVariable indexToRetrieve = indicesLoop.getView(SDIndex.point(index)).reshape(1).castTo("indexToReceive",DataType.INT64);
SDVariable indexToPut = indicesPut.getView(SDIndex.point(index)).reshape(1).castTo("indexToPut",DataType.INT64);
SDVariable toAssign = toPut.getView(SDIndex.point(indexToPut));
SDVariable sliceOutput = assignTo.getView(SDIndex.point(indexToRetrieve));
SDVariable assignOutput = loop.assign(sliceOutput,toAssign);
SDVariable outputIdentity = loop.identity("assignOutput",assignTo);
//ensure the output depends on the final assign so it gets executed, return the final output as a view
outputIdentity.addControlDependency(assignOutput);
return loop;
}
/**
* Create a graph that takes in the indices as a placeholder, loops over each element in the index vector
* and appends the slice to the end result. This graph is equivalent to something like:
* INDArray input = ....;
* INDArray indices = ...;
* INDArray result = input.get(NDArrayIndex.point(indices.getInt(0));
* for(int i = i; i < maxIndex && customInputResult; i++) {
* result = Nd4j.concat(0,input.get(NDArrayIndex.point(i)));
* }
* return result
*
* Note this is similar to {@link INDArray#get(INDArray)}
* @param relative the expected target input variable. We use this to pull expected
* return data type for the result
* @param indices the indices to get
* @return the graph for dynamically creating a result graph
*/
public static SameDiff createLoopConcat(SDVariable relative,SDVariable indices) {
//standard loop body for loopWithConditions
SameDiff loop = SameDiff.create();
//curr index
SDVariable index = loop.placeHolder("index",DataType.INT32);
//loop until
SDVariable maxIndex = loop.placeHolder("max",DataType.INT32);
//constant condition of true for custom, just loop till max iterations hit
SDVariable currCondition = loop.placeHolder("cond",DataType.BOOL);
//the input to pull from (in this case this)
SDVariable input = loop.placeHolder("input", relative.dataType());
//the actual variable to pull from
SDVariable pullFrom = loop.placeHolder("pullFrom",relative.dataType());
//the indices to loop over (the input variable
SDVariable indicesLoop = loop.placeHolder("indices",indices.dataType());
//standardize indices to length 1
indicesLoop = indicesLoop.reshape("indicesReshape",indicesLoop.length());
//the current index to retrieve
SDVariable indexToRetrieve = indicesLoop.get(SDIndex.point(index)).reshape(1).castTo("indexToReceive",DataType.INT64);
//the final concatenated output
SDVariable sliceOutput = loop.expandDims("outputSlice",pullFrom.get(SDIndex.point(indexToRetrieve)),0);
SDVariable output = loop.concat("output",0,input,sliceOutput);
return loop;
}
/**
* Convert this variable to a constant. This is equivalent to "freezing" a variable so that it's value
* won't be changed by further training.
* This can only be done for variables and placeholders, not ARRAY type variables (which are usually network activations).
* As a constant, this variable will no longer be modified by any subsequent training.
*
* @return This variable (now a constant)
*/
public SDVariable convertToConstant(){
return sameDiff.convertToConstant(this);
}
/**
* Convert this variable to a VARIABLE type SDVariable.
* This can only be done for constants and placeholders, not ARRAY type variables (which are usually network activations).
* As a variable, this variable will modified during any subsequent training.
*
* @return This variable (now a variable type SDVariable)
*/
public SDVariable convertToVariable(){
return sameDiff.convertToVariable(this);
}
/**
* Rename this variable to a new name. Equivalent to {@link SameDiff#renameVariable(String, String)}
*
* @param newName The new name for the variable - no variable with this name must already exist
* @return The current variable (same object)
*/
public SDVariable rename(String newName) {
sameDiff.renameVariable(getVarName(), newName);
return this;
}
/**
* Mark this variable as a loss function variable. This means that this variable will be minimized via backprop during training.
* This will add the variable as a loss to any others - i.e., if multiple variables are marked as losses, their values will be summed
* to give the total network loss.
* Note that only floating point (Float16/32/64) variables may be marked as a loss.
* Note also that only ARRAY type SDVariables can be marked as losses to be minimized. That is, we cannot mark the value
* of a constant, variable or placeholder to be minimized as doing so would not make sense.
* This is equivalent to {@link SameDiff#addLossVariable(String)}
*/
public void markAsLoss(){
sameDiff.addLossVariable(getVarName());
}
/**
* Determine if this variable has a gradient with respect to the current loss. Note that:
* (a) Non-floating-point variables (integer, string, etc) will never have gradients
* (b) This method will return false if no gradient function has been created yet. See {@link SameDiff#createGradFunction()}
* and {@link SameDiff#setLossVariables(String...)}
* (c) Floating point variables may not have any gradient if the current loss does not depend on the variable at all
* @return True if a gradient variable exists for the specified variable, for the current loss
*/
public boolean hasGradient(){
return sameDiff.variableHasGradient(getVarName());
}
private static int binArrToInt(int[] arr) {
int x = 0;
int m = 1;
for (int i = 0; i < arr.length; i++) {
if (arr[i] == 1) {
x += m;
}
m *= 2;
}
return x;
}
@Override
public int hashCode() {
int result = super.hashCode();
result = 31 * result + (varName != null ? varName.hashCode() : 0);
result = 31 * result + (variableType != null ? variableType.hashCode() : 0);
result = 31 * result + (dataType != null ? dataType.hashCode() : 0);
return result;
}
public SDVariable clone(String name,SameDiff sd) {
SDVariable v = new SDVariable();
v.varName = name;
v.variableType = variableType;
v.shape = shape == null ? null : shape.clone();
v.dataType = dataType;
v.sameDiff = sd;
return v;
}
public SDVariable clone(SameDiff sd) {
SDVariable v = new SDVariable();
v.varName = varName;
v.variableType = variableType;
v.shape = shape == null ? null : shape.clone();
v.dataType = dataType;
v.sameDiff = sd;
return v;
}
@Override
public boolean equals(Object o){
if(o == this) return true;
if(!(o instanceof SDVariable))
return false;
SDVariable s = (SDVariable)o;
if(!varName.equals(s.varName))
return false;
if(variableType != s.variableType)
return false;
if(dataType != s.dataType)
return false;
if(variableType == VariableType.VARIABLE || variableType == VariableType.CONSTANT){
INDArray a1 = getArr();
INDArray a2 = s.getArr();
return a1.equals(a2);
}
return true;
}
}